There is an AI model doing the round called Jev, from a company called TypeSafe AI, and I think it is interesting because it might be much closer to what corporations actually want from AI.
For the last few years, nearly all the public-facing excitement around AI has been about chatbots. You type something in, the AI thinks about it, and a big chunk of stuff comes back; it might be code, it might be power point, it might be plain text.
That is cool and everything, but it is not necessarily how you want to integrate AI into a corporate system.
Most enterprise applications do not need another person to talk to.
They need a decision that can then be ratified by a human or process.
Jev is what TypeSafe calls a System One Model. Rather than being designed primarily to generate text, it is designed to take some state, answer defined questions about that state, and return structured results that software can use directly.
TypeSafe describes this as “decisions, not strings”.
The model currently works around three basic types of answer:
- Noul: effectively a yes/no question returned as a probability.
- Choice: where the AI selects from a defined set of options.
- Score: where it returns a position on a defined scale.
The result is also accompanied by probability and confidence information.
This is much more interesting to me from an integration point of view.
Ive talked multiple times about how we actually implement AI into existing systems, with my favourite example being insurance fraud.
Imagine you already have a claims system.
Today it has a load of rules that say things such as:
If this condition is true, add five fraud points.
If that condition is true, add ten fraud points.
If the total goes above a certain level, send the claim for investigation.
AI can potentially make a much better judgement because it can look at far more context and far more subtle relationships.
But the person using the claims system does not need a chatbot appearing in the corner saying:
“Having carefully considered this claim, I believe there are several factors that may indicate…”
Nobody asked for an essay.
They need a dial or percentage.
Or simply “please investigate this claim”
That is a useful result.
The existing application and team can then do what it has always done but faster.
This is the bit I think we sometimes lose when talking about corporate AI.
AI should not necessarily be the application.
AI can simply be a component in an existing process, with the users just thinking its an improvement to the normal way they do things.
This is a very different architecture from throwing an autonomous chatbot at a business process and hoping it behaves itself.
TypeSafe’s own published workflow examples follow exactly this pattern. Their security incident example uses AI questions to assess an alert, then ordinary code determines whether to close it, queue it or take action. Their invoice example similarly combines AI judgements with deterministic business rules to decide whether something should be paid, held, disputed or sent for fraud review.
This is much closer to traditional enterprise integration.
And therefore, rather conveniently, it is much easier for corporate IT departments to understand.
Type-safe models do not mean magically correct; the Jev model can still make poor decisions.
Type-safe means, in this context, that you define the shape of the answer in advance.
If you ask it to choose between: Approve/Reject/Refer
You are not going to suddenly get three paragraphs discussing the philosophical nature of approval.
The software receives something it knows how to process.
That removes an entire class of integration rubbish.
- No trying to extract the answer from a paragraph.
- No checking whether it invented a fourth status.
- No defensive parsing because somebody changed the prompt.
The judgement might still be wrong, and you absolutely still need testing, monitoring and sensible confidence thresholds.
But the interface between the AI and the rest of your software becomes much more like an ordinary software interface.
Then there is the cost.
At the time of writing, TypeSafe lists Jev at $0.042 per million input tokens, with output tokens effectively free because it is not generating long strings of text. It also reports end-to-end response times of roughly 70 to 500 milliseconds.
Think how that applies to your big systems.
If a decision requires roughly 1,000 input tokens, the model cost at the published price is around $0.000042 per decision.
A million decisions of that size would therefore be around $42 in model input cost.
Obviously real systems have other costs: integration, infrastructure, monitoring, support, networking, data preparation, etc.
But the model cost suddenly becomes tiny for the right workload.
And there is another saving that is perhaps more important.
You are not paying a large general-purpose model to write an answer that your software then has to turn back into a usable decision.
TypeSafe reports very large speed and cost advantages over general-purpose LLMs on its own workflow evaluations 1.
Even ignoring the headline benchmark numbers, the savings make it worth looking at.
That, to me, is proper integration.
AI should become boring and just do the boring jobs for us.
Jev itself may or may not become the model everybody uses. It is extremely new, having been released by TypeSafe AI in September 2026, and real corporate implementations will tell us far more than launch demonstrations ever can.
But I think the architectural idea behind it is much more important.
Perhaps the future of corporate AI is not talking to it at all.
Perhaps it is just the thing quietly helping us make our decisions more accurate.
- although these are currently TypeSafe’s own benchmarks and so take with a pinch of salt[↩]